Human-in-the-Loop: When Agents Should Ask for Help (2026)

When to Use

• ✓Before irreversible actions: sending emails, deleting data, making payments, publishing content
• ✓When the agent's confidence score falls below a threshold (e.g., entity match confidence < 0.85)
• ✓For novel situations outside the agent's training distribution — detect via uncertainty signals
• ✓Compliance-required workflows where human sign-off is legally mandated
• ✓During initial deployment when you're still calibrating the agent's error rate before trusting autonomous operation

How It Works

1. 1Classify actions by reversibility and impact: reversible/low-impact (read, search, draft) = automate; reversible/high-impact (send draft for review) = optional confirmation; irreversible/any-impact (send final email, delete record, payment) = require confirmation.
2. 2Implement an interruption point as a special tool call: ask_human(question, context, options). The agent calls this tool when it needs input; execution pauses until the human responds.
3. 3For async workflows, queue the HITL request and notify the human via email/Slack. The agent state is serialized and resumed when the human responds. This requires durable execution infrastructure (Temporal, AWS Step Functions).
4. 4Show the human meaningful context — don't just ask 'Confirm action?'. Show: what action will be taken, why the agent decided this, what happens if the human says no, and what input is needed.
5. 5Track HITL interruptions over time. If a particular decision point triggers human review more than 5% of the time, it's a candidate for retraining or better prompt design. If it triggers less than 0.1%, consider removing the gate.

Examples

from anthropic import Anthropic

client = Anthropic()

tools = [
    {
        'name': 'send_email',
        'description': 'Send an email. ALWAYS call request_human_approval first for email sends.',
        'input_schema': {'type': 'object', 'properties': {'to': {'type': 'string'}, 'subject': {'type': 'string'}, 'body': {'type': 'string'}}, 'required': ['to', 'subject', 'body']}
    },
    {
        'name': 'request_human_approval',
        'description': 'Request human approval before performing a sensitive action. Call this before send_email, delete operations, and payments.',
        'input_schema': {'type': 'object', 'properties': {'action_description': {'type': 'string'}, 'action_details': {'type': 'object'}}, 'required': ['action_description', 'action_details']}
    }
]

# The model will call request_human_approval before send_email

def process_with_hitl(task: str, confidence_threshold: float = 0.85):
    result = agent.run(task)
    
    if result.confidence < confidence_threshold:
        # Agent is unsure — escalate
        return {
            'status': 'needs_review',
            'draft': result.output,
            'confidence': result.confidence,
            'uncertain_aspects': result.low_confidence_parts,
            'reviewer_url': create_review_task(result)
        }
    
    return {'status': 'complete', 'output': result.output}

Common Mistakes

• ✗HITL for everything — if the agent asks for human confirmation on every step, users bypass it entirely or abandon the system. Reserve HITL for genuinely high-stakes decisions. Start with broad HITL and narrow it as you build confidence.
• ✗Not providing enough context in HITL requests — 'Do you want to proceed?' with no context forces the human to dig through logs. Show: action, reasoning, impact, and alternatives. The human should be able to approve/reject in under 30 seconds.
• ✗Synchronous HITL that blocks the system — for async workflows, HITL requests should not block other work. Queue the request, serialize agent state, and resume when the human responds. Synchronous HITL that blocks a thread is a scalability problem.
• ✗No timeout handling for HITL requests — if the human doesn't respond in 24 hours, the agent should either escalate or take a safe default action. Undefined timeouts cause agent tasks to hang indefinitely.

FAQ